Related Experiment Video
Updated: Sep 12, 2025
04:48
Swin-PSAxialNet: An Efficient Multi-Organ Segmentation Technique
Published on: July 5, 2024
493
LKDA-Net: Hierarchical transformer with large Kernel depthwise convolution attention for 3D medical image
Ming Li1, Jingang Ma2, Jing Zhao3
1Graduate School, Shandong University of Traditional Chinese Medicine, Jinan, China.
Plos One
|August 8, 2025
Summary
LKDA-Net, a lightweight 3D convolutional network, achieves efficient and accurate medical image segmentation. It uses large-kernel depthwise convolutions to mimic Transformer self-attention, outperforming existing models with fewer parameters.
Area of Science:
- Medical Image Analysis
- Computer Vision
- Deep Learning
Background:
- Transformers excel in 2D medical image segmentation and are being adapted for 3D tasks.
- Hierarchical transformers like Swin UNETR integrate convolutional network priors for enhanced 3D segmentation.
- Current hybrid methods rely on numerous parameters and large receptive fields from self-attention.
Purpose of the Study:
- To propose a lightweight 3D convolutional network, LKDA-Net, for efficient and accurate 3D volumetric segmentation.
- To explore large-kernel depthwise convolutions as an alternative to self-attention for achieving large receptive fields with fewer parameters.
- To enhance feature expression and segmentation performance through novel architectural components.
Main Methods:
- Developed LKDA-Net, a lightweight 3D convolutional network utilizing a large-kernel depthwise convolution attention mechanism.
- Investigated various large-kernel sizes inspired by Swin Transformer modules to achieve global receptive fields.
- Replaced the MLP in Swin Transformer with an Inverted Bottleneck with Depthwise Convolutional Augmentation for improved feature representation.
- Introduced a skip connection fusion module for effective encoder-decoder feature fusion.
Main Results:
- LKDA-Net demonstrated superior segmentation performance compared to existing models across three public datasets (Synapse, BTCV, ACDC).
- The proposed network achieved this performance with a significantly lower parameter count.
- The large-kernel depthwise convolution attention effectively simulated Transformer self-attention benefits.
Conclusions:
- LKDA-Net offers an efficient and accurate solution for 3D medical volumetric segmentation.
- The network's design, particularly the use of large-kernel depthwise convolutions, provides a parameter-efficient alternative to Transformer-based methods.
- LKDA-Net shows strong potential for clinical applications requiring high-performance medical image segmentation.
